lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Portal-RPC reconnection and replay operations, for use in recovery.
   5  *
   6  *  Copyright (c) 2003 Cluster File Systems, Inc.
   7  *   Authors: Phil Schwan <phil@clusterfs.com>
   8  *            Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef __KERNEL__
  27 #include <liblustre.h>
  28 #else
  29 #include <linux/version.h>
  30 #include <asm/semaphore.h>
  31 #define DEBUG_SUBSYSTEM S_RPC
  32 #endif
  33
  34 #include <linux/obd_support.h>
  35 #include <linux/obd_class.h>
  36 #include "ptlrpc_internal.h"
  37
  38 static DECLARE_MUTEX(pinger_sem);
  39 static struct list_head pinger_imports = LIST_HEAD_INIT(pinger_imports);
  40
  41 static struct ptlrpc_thread *pinger_thread = NULL;
  42
  43 int ptlrpc_ping(struct obd_import *imp)
  44 {
  45         struct ptlrpc_request *req;
  46         int rc = 0;
  47         ENTRY;
  48
  49         req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL,
  50                               NULL);
  51         if (req) {
  52                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
  53                           imp->imp_obd->obd_uuid.uuid,
  54                           imp->imp_target_uuid.uuid);
  55                 req->rq_no_resend = req->rq_no_delay = 1;
  56                 req->rq_replen = lustre_msg_size(0,
  57                                                  NULL);
  58                 ptlrpcd_add_req(req);
  59         } else {
  60                 CERROR("OOM trying to ping %s->%s\n",
  61                           imp->imp_obd->obd_uuid.uuid,
  62                           imp->imp_target_uuid.uuid);
  63                 rc = -ENOMEM;
  64         }
  65
  66         RETURN(rc);
  67 }
  68
  69 #ifdef __KERNEL__
  70 static int ptlrpc_pinger_main(void *arg)
  71 {
  72         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
  73         struct ptlrpc_thread *thread = data->thread;
  74         unsigned long flags;
  75         ENTRY;
  76
  77         lock_kernel();
  78         ptlrpc_daemonize();
  79
  80         SIGNAL_MASK_LOCK(current, flags);
  81         sigfillset(&current->blocked);
  82         RECALC_SIGPENDING;
  83         SIGNAL_MASK_UNLOCK(current, flags);
  84
  85         THREAD_NAME(current->comm, "%s", data->name);
  86         unlock_kernel();
  87
  88         /* Record that the thread is running */
  89         thread->t_flags = SVC_RUNNING;
  90         wake_up(&thread->t_ctl_waitq);
  91
  92         /* And now, loop forever, pinging as needed. */
  93         while (1) {
  94                 unsigned long this_ping = jiffies;
  95                 long time_to_next_ping;
  96                 struct l_wait_info lwi = LWI_TIMEOUT(obd_timeout * HZ,
  97                                                      NULL, NULL);
  98                 struct list_head *iter;
  99
 100                 down(&pinger_sem);
 101                 list_for_each(iter, &pinger_imports) {
 102                         struct obd_import *imp =
 103                                 list_entry(iter, struct obd_import,
 104                                            imp_pinger_chain);
 105                         int force, level;
 106                         unsigned long flags;
 107
 108
 109                         spin_lock_irqsave(&imp->imp_lock, flags);
 110                         level = imp->imp_state;
 111                         force = imp->imp_force_verify;
 112                         if (force)
 113                                 imp->imp_force_verify = 0;
 114                         spin_unlock_irqrestore(&imp->imp_lock, flags);
 115
 116                         if (imp->imp_next_ping <= this_ping || force) {
 117                                 if (level == LUSTRE_IMP_DISCON) {
 118                                         /* wait at least a timeout before
 119                                            trying recovery again. */
 120                                         imp->imp_next_ping = jiffies +
 121                                                 (obd_timeout * HZ);
 122                                         ptlrpc_initiate_recovery(imp);
 123                                 }
 124                                 else if (level != LUSTRE_IMP_FULL ||
 125                                          imp->imp_obd->obd_no_recov) {
 126                                         CDEBUG(D_HA,
 127                                                "not pinging %s (in recovery "
 128                                                " or recovery disabled: %s)\n",
 129                                                imp->imp_target_uuid.uuid,
 130                                                ptlrpc_import_state_name(level));
 131                                 }
 132                                 else if (imp->imp_pingable || force) {
 133                                         ptlrpc_ping(imp);
 134                                 }
 135
 136                         } else {
 137                                 if (imp->imp_pingable)
 138                                         CDEBUG(D_HA, "don't need to ping %s "
 139                                                "(%lu > %lu)\n",
 140                                                imp->imp_target_uuid.uuid,
 141                                                imp->imp_next_ping, this_ping);
 142                         }
 143                 }
 144                 up(&pinger_sem);
 145
 146                 /* Wait until the next ping time, or until we're stopped. */
 147                 time_to_next_ping = this_ping + (obd_timeout * HZ) - jiffies;
 148                 CDEBUG(D_HA, "next ping in %lu (%lu)\n", time_to_next_ping,
 149                        this_ping + (obd_timeout * HZ));
 150                 if (time_to_next_ping > 0) {
 151                         lwi = LWI_TIMEOUT(time_to_next_ping, NULL, NULL);
 152                         l_wait_event(thread->t_ctl_waitq,
 153                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 154                                      &lwi);
 155                         if (thread->t_flags & SVC_STOPPING) {
 156                                 thread->t_flags &= ~SVC_STOPPING;
 157                                 EXIT;
 158                                 break;
 159                         } else if (thread->t_flags & SVC_EVENT) {
 160                                 /* woken after adding import to reset timer */
 161                                 thread->t_flags &= ~SVC_EVENT;
 162                         }
 163                 }
 164         }
 165
 166         thread->t_flags = SVC_STOPPED;
 167         wake_up(&thread->t_ctl_waitq);
 168
 169         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current->pid);
 170         return 0;
 171 }
 172
 173 int ptlrpc_start_pinger(void)
 174 {
 175         struct l_wait_info lwi = { 0 };
 176         struct ptlrpc_svc_data d;
 177         int rc;
 178 #ifndef ENABLE_PINGER
 179         return 0;
 180 #endif
 181         ENTRY;
 182
 183         if (pinger_thread != NULL)
 184                 RETURN(-EALREADY);
 185
 186         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 187         if (pinger_thread == NULL)
 188                 RETURN(-ENOMEM);
 189         init_waitqueue_head(&pinger_thread->t_ctl_waitq);
 190
 191         d.name = "ll_ping";
 192         d.thread = pinger_thread;
 193
 194         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 195          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 196         rc = kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 197         if (rc < 0) {
 198                 CERROR("cannot start thread: %d\n", rc);
 199                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 200                 RETURN(rc);
 201         }
 202         l_wait_event(pinger_thread->t_ctl_waitq,
 203                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 204
 205         RETURN(rc);
 206 }
 207
 208 int ptlrpc_stop_pinger(void)
 209 {
 210         struct l_wait_info lwi = { 0 };
 211         int rc = 0;
 212 #ifndef ENABLE_PINGER
 213         return 0;
 214 #endif
 215         ENTRY;
 216
 217         if (pinger_thread == NULL)
 218                 RETURN(-EALREADY);
 219         down(&pinger_sem);
 220         pinger_thread->t_flags = SVC_STOPPING;
 221         wake_up(&pinger_thread->t_ctl_waitq);
 222         up(&pinger_sem);
 223
 224         l_wait_event(pinger_thread->t_ctl_waitq,
 225                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 226
 227         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 228         pinger_thread = NULL;
 229         RETURN(rc);
 230 }
 231
 232 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 233 {
 234         down(&pinger_sem);
 235         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 236         up(&pinger_sem);
 237 }
 238
 239 int ptlrpc_pinger_add_import(struct obd_import *imp)
 240 {
 241         ENTRY;
 242         if (!list_empty(&imp->imp_pinger_chain))
 243                 RETURN(-EALREADY);
 244
 245         down(&pinger_sem);
 246         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 247                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 248         imp->imp_next_ping = jiffies + (obd_timeout * HZ);
 249         /* XXX sort, blah blah */
 250         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 251         class_import_get(imp);
 252
 253         ptlrpc_pinger_wake_up();
 254         up(&pinger_sem);
 255
 256         RETURN(0);
 257 }
 258
 259 int ptlrpc_pinger_del_import(struct obd_import *imp)
 260 {
 261         ENTRY;
 262         if (list_empty(&imp->imp_pinger_chain))
 263                 RETURN(-ENOENT);
 264
 265         down(&pinger_sem);
 266         list_del_init(&imp->imp_pinger_chain);
 267         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 268                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 269         class_import_put(imp);
 270         up(&pinger_sem);
 271         RETURN(0);
 272 }
 273
 274 void ptlrpc_pinger_wake_up()
 275 {
 276 #ifdef ENABLE_PINGER
 277         pinger_thread->t_flags |= SVC_EVENT;
 278         wake_up(&pinger_thread->t_ctl_waitq);
 279 #endif
 280 }
 281
 282 #else /* !__KERNEL__ */
 283
 284 /* XXX
 285  * the current implementation of pinger in liblustre is not optimized
 286  */
 287
 288 static struct pinger_data {
 289         int             pd_recursion;
 290         unsigned long   pd_this_ping;
 291         unsigned long   pd_next_ping;
 292         int             pd_force_check;
 293 } pinger_args;
 294
 295 static int pinger_check_rpcs(void *arg)
 296 {
 297         unsigned long curtime = time(NULL);
 298         struct list_head *iter;
 299         struct pinger_data *pd = &pinger_args;
 300
 301         /* prevent recursion */
 302         if (pd->pd_recursion++) {
 303                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 304                 pd->pd_recursion--;
 305                 return 0;
 306         }
 307
 308         /* have we reached ping point? */
 309         if (pd->pd_next_ping > curtime && !pd->pd_force_check) {
 310                 pd->pd_recursion--;
 311                 return 0;
 312         }
 313
 314         if (pd->pd_force_check)
 315                 pd->pd_force_check = 0;
 316
 317         pd->pd_this_ping = curtime;
 318
 319         /* add rpcs into set */
 320         down(&pinger_sem);
 321         list_for_each(iter, &pinger_imports) {
 322                 struct obd_import *imp =
 323                         list_entry(iter, struct obd_import,
 324                                    imp_pinger_chain);
 325                 int level, force;
 326                 unsigned long flags;
 327
 328
 329                 spin_lock_irqsave(&imp->imp_lock, flags);
 330                 level = imp->imp_state;
 331                 force = imp->imp_force_verify;
 332                 if (force)
 333                         imp->imp_force_verify = 0;
 334                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 335
 336                 if (imp->imp_next_ping <= pd->pd_this_ping || force) {
 337                         if (level == LUSTRE_IMP_DISCON) {
 338                                 /* wait at least a timeout before
 339                                    trying recovery again. */
 340                                 unsigned long timeout = obd_timeout;
 341                                 if (imp->imp_server_timeout)
 342                                         timeout = obd_timeout / 2;
 343                                 imp->imp_next_ping = time(NULL) +
 344                                         (timeout * HZ);
 345                                 ptlrpc_initiate_recovery(imp);
 346                         }
 347                         else if (level != LUSTRE_IMP_FULL ||
 348                                  imp->imp_obd->obd_no_recov) {
 349                                 CDEBUG(D_HA,
 350                                        "not pinging %s (in recovery "
 351                                        " or recovery disabled: %s)\n",
 352                                        imp->imp_target_uuid.uuid,
 353                                        ptlrpc_import_state_name(level));
 354                         }
 355                         else if (imp->imp_pingable || force) {
 356                                 ptlrpc_ping(imp);
 357                         }
 358
 359                 } else {
 360                         if (imp->imp_pingable) {
 361                                 CDEBUG(D_HA, "don't need to ping %s "
 362                                        "(%lu > %lu)\n",
 363                                        imp->imp_target_uuid.uuid,
 364                                        imp->imp_next_ping, pd->pd_this_ping);
 365                         }
 366                 }
 367         }
 368
 369         up(&pinger_sem);
 370
 371         pd->pd_next_ping = pd->pd_this_ping + (obd_timeout * HZ);
 372
 373         CDEBUG(D_HA, "finished a round ping\n");
 374         pd->pd_recursion--;
 375         return 0;
 376 }
 377
 378 static void *pinger_callback = NULL;
 379
 380 int ptlrpc_start_pinger(void)
 381 {
 382         memset(&pinger_args, 0, sizeof(pinger_args));
 383 #ifdef ENABLE_PINGER
 384         pinger_callback =
 385                 liblustre_register_wait_callback(&pinger_check_rpcs, &pinger_args);
 386 #endif
 387         return 0;
 388 }
 389
 390 int ptlrpc_stop_pinger(void)
 391 {
 392 #ifdef ENABLE_PINGER
 393         if (pinger_callback)
 394                 liblustre_deregister_wait_callback(pinger_callback);
 395 #endif
 396         return 0;
 397 }
 398
 399 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 400 {
 401         down(&pinger_sem);
 402         imp->imp_next_ping = time(NULL) + obd_timeout;
 403         if (pinger_args.pd_next_ping > imp->imp_next_ping) {
 404                 CDEBUG(D_HA, "set next ping to %ld(cur %ld)\n",
 405                         imp->imp_next_ping, time(NULL));
 406                 pinger_args.pd_next_ping = imp->imp_next_ping;
 407         }
 408         up(&pinger_sem);
 409 }
 410
 411 int ptlrpc_pinger_add_import(struct obd_import *imp)
 412 {
 413         ENTRY;
 414         if (!list_empty(&imp->imp_pinger_chain))
 415                 RETURN(-EALREADY);
 416
 417         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 418                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 419         ptlrpc_pinger_sending_on_import(imp);
 420
 421         down(&pinger_sem);
 422         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 423         class_import_get(imp);
 424         up(&pinger_sem);
 425
 426         RETURN(0);
 427 }
 428
 429 int ptlrpc_pinger_del_import(struct obd_import *imp)
 430 {
 431         ENTRY;
 432         if (list_empty(&imp->imp_pinger_chain))
 433                 RETURN(-ENOENT);
 434
 435         down(&pinger_sem);
 436         list_del_init(&imp->imp_pinger_chain);
 437         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 438                imp->imp_obd->obd_uuid.uuid, imp->imp_target_uuid.uuid);
 439         class_import_put(imp);
 440         up(&pinger_sem);
 441         RETURN(0);
 442 }
 443
 444 void ptlrpc_pinger_wake_up()
 445 {
 446         pinger_args.pd_force_check = 1;
 447 }
 448 #endif /* !__KERNEL__ */